Correction (19 October 2026): An earlier version of this article described events that we could not verify. It has been rewritten from verifiable sources, which are linked in the text.

On 31 July 2026 the 42nd Civil Chamber of Munich Regional Court I (Landgericht München I) largely upheld the claims of GEMA, Germany's music collecting society, against Suno, the US company behind a popular AI music generator. According to the court's press release, the chamber granted GEMA's claims for an injunction, for information and for damages in substance (case no. 42 O 763/25). It held that the way six songs were memorised in Suno's models, and then resurfaced in its outputs, infringed German copyright. This is a first-instance judgment and it is not final.

The earlier version of this article said that Germany's Federal Court of Justice had ruled on 3 August 2026, that Udio was also a defendant, that the court had extended protection to an artist's "voice timbre" and "style", that it had ordered an extended collective licensing scheme, and that it had required AI music to be labelled and royalty pools to be separated. None of this appears in the court's press release or in the reporting we reviewed, and all of it has been removed.

What the Munich court decided

The parties and the six works

GEMA filed the action on 21 January 2025. In GEMA's account, Suno did not respond to repeated requests to license the works, and while it admitted training on GEMA repertoire without paying for it, it disputed any obligation to pay. The case covered six titles: "Atemlos durch die Nacht", "Rasputin", "Big in Japan", "Forever Young", the refrain of "Mambo No. 5" and "Daddy Cool". The press release notes that the lyrics were not at issue. The dispute was about the music itself. GEMA's evidence rested on outputs generated with simple prompts that gave each song's original lyrics, its title and a desired style, but nothing on melody, harmony, rhythm or arrangement. According to GEMA, those outputs closely matched the originals in melody, harmony and rhythm (press release; GEMA; t3n).

Memorisation in the model is reproduction

Suno argued that its models contain no copies of songs, only generalised patterns learned from large datasets, and that any similarities came from the prompts or from statistical regularities. The chamber disagreed. It found that the six works were reproducible from Suno's model versions v3.5 and v4, which were stored on servers in Germany. It reached that conclusion by comparing the training works with the outputs, and given the complexity and length of the works it ruled out coincidence. That memorisation infringed the reproduction right under § 16 of the German Copyright Act (UrhG). The court also held that the text and data mining exception in § 44b UrhG did not cover it (press release; miz).

Outputs, and who answers for them

The recognisable original elements in outputs generated in Germany were held to be unauthorised reproductions and communications to the public. The court placed responsibility on Suno, not on its users, because Suno operates the models and chose the training data. It also held that offering the model and the generation service infringed the unnamed right of communication to the public in § 15(2) UrhG (miz; t-online).

Training in the United States

Suno's training took place in the US. Under the principle that the law of the country for which protection is claimed applies, the court assessed that conduct under US law and held that it was not fair use under 17 U.S.C. § 107. It distinguished the US decisions in Bartz and Kadrey, where training was found to be fair use because the outputs did not substantially reproduce the training works. In Suno's case they did (press release; miz). According to t-online, the court also pointed to the circumvention of download protection and to the commercial character of the use. The facts summarised in the press release state that Suno had used stream-ripping techniques to extract the works from YouTube, circumventing the platform's "rolling cipher", a technical measure designed to prevent downloads. The court based its international jurisdiction over the US-related claims on § 131 of the Collecting Societies Act (VGG), which gives collecting societies a special venue.

Remedies, reactions and status

Suno must stop the infringing reproductions, provide information and pay damages. The amount of damages has not yet been set (Music Week). According to t-online, Suno succeeded only on minor side issues. The judgment is not final. A Suno spokesperson said, "We disagree with today's ruling." The company said the decision rests on a mischaracterisation of how its technology works and how US law applies, and that it was evaluating all options, including an appeal (Music Week; ZDF). As of 19 October 2026 we have not found a reliable report confirming whether an appeal has been lodged.

GEMA's chief executive, Tobias Holzmüller, said AI providers must take licences rather than use members' works free of charge (GEMA). GEMA says it seeks payment through licensing, not a general ban on AI training (ZDF). Wolfram Weimer, Germany's Minister of State for Culture, said the ruling showed that the protection of human creative work is indispensable and applies to providers of generative AI too. He called for a framework that rewards innovation while paying rights holders fairly (t-online).

The line from GEMA v. OpenAI

The Suno judgment builds on the same chamber's decision of 11 November 2025 in GEMA v. OpenAI (42 O 14139/24). That case was filed on 13 November 2024 and concerned the lyrics of nine well-known German songs, among them "Atemlos" and "Wie schön, dass du geboren bist" (MediaLaws; epd/evangelisch.de). The court found that the lyrics were memorised in OpenAI's language models. It held that text and data mining exists to extract information, and that memorising whole works goes beyond that, so neither § 44b nor § 60d UrhG applied. The outputs were unauthorised communications to the public, the pastiche exception was rejected, and OpenAI was held directly responsible rather than treated as a neutral intermediary. Injunctions, information and damages were granted (MediaLaws). OpenAI has appealed to the Higher Regional Court of Munich (GEMA; ZDF).

Suno extends that reasoning from text to music, and to a model trained in the US but stored and offered in Germany.

Courts do not all see it this way. In Getty Images v Stability AI [2025] EWHC 2863 (Ch), the English High Court held that the Stable Diffusion model weights were not "infringing copies" because they did not store the training images. Getty had dropped its training claim because there was no evidence that training took place in the UK (Faegre Drinker). The cases differ in statute, medium and evidence.

Where EU law stands

The Munich decisions are national first-instance rulings. The questions they answer are now before the Court of Justice of the EU in Like Company v Google Ireland (C-250/25), referred by a Budapest court on 3 April 2025. The referring court asks whether training a model is reproduction, whether the text and data mining exception in Article 4 of the DSM Directive covers it, and whether outputs that reproduce protected content amount to communication to the public attributable to the provider (Stephenson Harwood). ZDF noted that some of these questions may ultimately have to be settled at EU level and that a final outcome could take years.

The EU AI Act adds separate obligations that do not depend on any court ruling. Since 2 August 2025, providers of general-purpose AI models must have a policy to comply with EU copyright law, including respecting rights reservations made under Article 4(3) of the DSM Directive, and must publish a sufficiently detailed summary of the content used for training (Article 53(1)(c) and (d)). Whether a dedicated music generator counts as a general-purpose model is a separate question. Under Article 50(2), providers of generative AI systems must mark synthetic audio, images, video and text in a machine-readable, detectable format (William Fry). The AI Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026. It gives systems placed on the market before 2 August 2026 until 2 December 2026 to comply with that marking duty. New systems are bound from 2 August 2026 (Cooley). Any labelling duty for AI music in the EU therefore comes from the AI Act, not from the Munich judgment.

Türkiye

Turkish law has no court decision comparable to the Munich judgments that we could find. Its starting point is wide, however. Article 22 of the Law on Intellectual and Artistic Works (FSEK, No. 5846) gives the author the exclusive right to reproduce the original or copies of a work in any form or by any method, wholly or partly, directly or indirectly, temporarily or permanently. Recording a work on any device that transmits or repeats sound or images also counts as reproduction (FSEK). Unlike the EU, FSEK has no text and data mining exception, so only its existing statutory exceptions are available. Authorship is limited to natural persons, and pending legislative proposals reportedly envisage a licensing-based regime with remuneration for AI training (Chambers, Türkiye chapter by KM Legal). We have not been able to verify the text of those proposals. FSEK does not expressly address automated data processing (İstanbul 29 Mayıs University law blog).

Korea

Korea regulates transparency first. The Framework Act on AI took effect on 22 January 2026. Its Article 31 requires providers to tell users in advance when a product uses generative AI and to indicate that outputs were AI-generated. Realistic synthetic sound, images or video must be disclosed in a way users can clearly recognise. Failure to give notice can lead to a fine of up to KRW 30 million (Future of Privacy Forum). On copyright, the Ministry of Culture, Sports and Tourism's guide advises AI operators to secure legitimate rights to training data with appropriate compensation and to prevent outputs identical or similar to existing works. It also states that only works expressing human creativity can be registered (MCST). Two further guides from June 2025 cover registration of AI-assisted works and the prevention of disputes over outputs (Mondaq). The government's AI Action Plan proposes legal changes so that works can be used for AI training "without legal uncertainty". In January 2026, sixteen creators' organisations, including music rights holders, rejected this as a "use first, pay later" approach (Korea Herald via Asia News Network). We could not confirm whether an amendment has since been tabled or passed.

Burhan Doğuş Ayparlar's View

In our view, the real Munich judgment is narrower than the earlier version of this article suggested, and on that account more persuasive. It does not protect a voice or a style. It rests on evidence that particular, identifiable works can be drawn out of a particular model with simple prompts. That method can be repeated by any rights holder, and it shifts the burden to design. If a provider's model reproduces recognisable works on ordinary prompts, the chamber treats the provider, not the user, as the one who reproduces and communicates them.

The cross-border reasoning is the most consequential part. If it holds on appeal, training in a fair-use jurisdiction will not shield a provider whose model is stored and offered in Germany, and a collecting society can bring the US-law claims before a German court. That is a large "if": the OpenAI appeal is pending in Munich, and the CJEU's answers in C-250/25 may reshape the analysis across the EU. The reasoning also turns on memorisation. Training copies that never resurface in a model's outputs raise a different question, one the Munich decisions do not settle.

For Türkiye, the absence of a text and data mining exception cuts both ways. Rights holders do not have to show an effective opt-out, but providers have no statutory safe harbour to point to, and the breadth of Article 22 FSEK leaves little doubt that a fixed copy is a reproduction. What is missing is proof and practice: we know of no Turkish case that has tested a model's memorisation. Imitation of a singer's voice, as distinct from copying a recording or a composition, is not a copyright question in the Munich sense. In Türkiye it is more naturally argued under personality rights and personal data law.

Sources

This article is general information and does not constitute legal advice. It reflects the position as of 19 October 2026.